In the normalizing flows section, a MAF is defined with relu activation. Wouldn't this violate the bijectivity of the model?